#!/bin/bash

# Eval Qwen3-VL CoT reranker with generation and final-answer parsing
CUDA_VISIBLE_DEVICES=3 python eval/eval_reranker_qwen3vl_cot.py \
    --base-model /data/oss_bucket_0/liqi/models/Qwen3-VL-2B-Instruct/ \
    --model-path /data/oss_bucket_0/liqi/cot_reranker/cot_sft/qwen3-vl-2b-cot-reranker_lora_True/checkpoint-860/ \
    --eval-data /data/oss_bucket_0/liqi/dataset/reranker/test.jsonl \
    --no-full-crop-mode \
    --format-prompt train_cot/listwise_cot_reranker_v1.jinja \
    --model-max-length 10240 \
    --max-new-tokens 2048 \
    --max-crop-tokens 80 \
    --max-full-tokens 20 \
    --output-dir /data/oss_bucket_0/liqi/cot_reranker/cot_sft/qwen3-vl-2b-cot-reranker_lora_True/eval_results